LLM-Driven Optical Network-as-a-Service (ONaaS) for Next Generation Metro-Access Convergence

The convergence of metro and access domains and the shift towards intent-driven, autonomous optical infrastructures are accelerating the demand for intelligent control systems. This paper investigates the feasibility of deploying Large Language Models (LLMs) as agents to support an Optical-Networkas- a-Service (ONaaS) paradigm for next-generation metro–access convergence. We introduce an LLM-driven framework that integrates natural-language intent interpretation, topology generation, and network equipment allocation into a unified agent architecture. Eight state-of-the-art open-source LLMs, including LLaMA-3, Mistral-7b, Gemma-2/3, and Qwen-2.5, are benchmarked under Zero-Shot, One-Shot, and Few-Shot prompting strategies using a standardized evaluation criterion based on tool-call validity, hallucination rate, and inference latency. The results show that example-driven prompting improves reliability for most models, with the strongest and most consistent gains observed for Mistral-7b and LLaMA-3-8b, while some families (e.g., Gemma and Qwen variants) showed limited or mixed improvements under Few-Shot prompts. Mistral-7b achieved the highest success rate, however, with higher inference latency, whereas LLaMA-3-8b offers the best success-to-latency tradeoff. Overall, our ONaaS scenarios demonstrate that open-source LLMs can robustly orchestrate network planning functions from natural-language intents, paving the way for the proposed architecture to integrate designed impairment-aware modules in future extensions, thereby enabling tighter integration with digital-twin environments and impairment-aware planning workflows.